WEKA

AI Infrastructure · Data Storage · Enterprise IT

Launch

WEKA debuts NeuralMesh 6 and WEKApod 3, its first custom AI storage hardware

Extending GPU memory with cheap flash could let AI clouds serve far more tokens without buying more chips.

• WEKA unveiled NeuralMesh 6, a revamped AI data and memory software platform, alongside WEKApod 3 — the company's first self-designed hardware line — built for agentic AI inference at production scale.
• WEKA (formerly WekaIO) is an AI data and memory infrastructure company whose Augmented Memory Grid extends GPU memory by caching persistent KV-cache tokens on NVMe flash instead of pricier GPU RAM.
• A single WEKApod 3 rack delivers 1.1 exabytes of effective capacity from 441.5PB raw storage, 10.2TB/s throughput and 210 million IOPS — the first single-rack system to break the exabyte barrier.
• Benchmarks on Oracle Cloud Infrastructure H100 systems showed NeuralMesh's Augmented Memory Grid delivering 10x higher token throughput, 10x more concurrent users, and 7x more tokens per GPU.
• The launch lands in an increasingly crowded field, as Dell, NetApp, Pure Storage and VAST have all repositioned their storage lines toward AI infrastructure over the past two years.
• As inference, not training, becomes the cost bottleneck for AI companies, treating cheap flash as extended GPU memory offers a way to scale agentic workloads without endlessly buying more chips.